基于改进mRMR特征选择的云型识别研究

Study of Cloud-Type Recognition Based on an Improved mRMR Feature Selection Method

传统的云型识别主要是提取云的颜色、纹理和形状等特征,但这些特征中存在不相关和冗余特征,导致云型识别率降低。在最大相关最小冗余(max-relevance and min-redundancy,mRMR)特征选择方法的基础上,运用互信息标准化形式(Symmetrical Uncertainty,SU)克服互信息偏向于取值较多属性的固有缺点,提出了改进的mRMR特征选择方法,对云的综合特征集进行特征筛选,筛选出最优特征子集,运用支持向量机进行云型识别。试验结果表明该方法优于mRMR方法,使层云、积云、高积云、卷云和晴空5种天空类型的总正确率提高,特征选择前、后的总识别率分别为86.96%、89.04%,识别率提高了2%;对于云型识别研究,经过特征选择后可知纹理特征优于形状特征,基于形状的Zernike矩优于HU不变矩,基于纹理的灰度共生矩阵为最优特征提取方法。

In the traditional cloud-type recognition method, a set of features describing the color, texture and shape features of clouds are extracted, in which there are some irrelevance and redundancy features leading to the reduced recognition rate of cloud-type. Based on the criteria of the max-relevance and min- redundancy （mRMR）, symmetrical uncertainty is employed to overcome the inherent defect of mutual information, which tends to have more value attributes. The improved mRMR feature selection method is putted forward, and the best feature subsets are selected by this method, and then the support vector machine is used to the recognition of cloud-type： Experimental results show that the correct recognition rate of altocumulus, cirrus, clear, cumulus, and stratus are improved significantly, with the total recognition rate being 86.96%; after feature selection, the total recognition rate can increase to 89.04%, and the recognition rate increases by 2%. For cloud type classification research, the texture feature is better than the shape feature; the shape features based on Zernike moment is better than HU moment invariants; the texture feature based on the gray level co-occurrence matrix is the optimum feature extraction method.